| { |
| "version": "question_conditioned_answer_adjudication.v2.gap_a8", |
| "student_outputs_used": false, |
| "source_gold_modified": false, |
| "rules": { |
| "explicit_onset": "Correct onset within 0.1s; no required end.", |
| "generic_when_with_selected_occurrence": "A correct onset (including for-the-first-time) or a correct full interval is acceptable. Do not infer a required end solely from A1 or the gold interval.", |
| "generic_when_multiple_occurrences_without_selector": "Cover all relevant occurrences, as onsets or intervals. A single first occurrence is insufficient.", |
| "explicit_time_range": "When the user asks for the span/from-until/start-and-end, require both endpoints.", |
| "duration": "Require the requested instance/total/max duration. Instance selection must be supported by wording/history.", |
| "optional_detail": "Accurate unrequested detail is harmless; contradictory detail remains incorrect.", |
| "span_evidence": "Dense occurrence evidence for ordinary tasks; query-conditioned derived gap interval for A5-gap, with supporting source occurrences stored separately. Complete Route, interval count, NONE and legality still required.", |
| "gap": "Resolve the endpoint occurrences from question/history before using gold. Positive gap=[end(left),start(right)]. Overlap/touching is not a positive-length gap. A gap relative to an event does not imply globally silent audio. Round endpoints independently to 0.1 s, then subtract for canonical duration; preserve raw endpoints/difference and allow 0.1 s numeric tolerance.", |
| "A8": "Count complete separate target occurrences over all explicitly requested clips, with zero for verified annotation absence. Return every clip tied at the maximum; counts/times are optional, contradictory additions remain incorrect." |
| }, |
| "reason": "Operational clarification of initial question-conditioned protocol before any model test; avoids inventing an unasked end-time requirement.", |
| "parent_version": "question_conditioned_answer_adjudication.v1", |
| "source_original_turns_modified": false, |
| "curator_authored_turns": 8 |
| } |
|
|
